Tops GPTs 
PrismML adapted its 1-bit Bonsai LLM to run locally on smart glasses using Qualcomm’s Snapdragon AR1 Gen 1 Platform, the company says. The 2-billion-parameter model is tuned for vision-and-language and, PrismML claims, is a 4x-shrunk version that keeps near-original benchmark performance.
The Transformers team added support for running GGUF (llama.cpp) quantized checkpoints inside transformers, with an initial focus on Apple Silicon and Qwen3.5. Benchmarks reported by the team show performance close to llama.cpp on a MacBook Pro M2 Max.
OpenAI says GPT‑6 now uses an improved prompt caching system that raises cache hit rates and offers discounts of up to 90% on cached input tokens. New tools include the Prompt Caching Dashboard, a diagnostics tool, explicit cache controls and prewarming.
OpenAI introduced GPT‑6 Sol and GPT‑6 Luna as lower-cost variants trained with methods from GPT‑6 Astra. The company reports up to 50% lower API prices versus GPT‑5.6 and presents benchmark and caching improvements it says make the models more cost‑efficient.
At its Made On YouTube event, YouTube introduced AI features in YouTube Studio to help creators analyze drafts, generate thumbnails and titles, and study trending formats. The app will expand Ask Studio to iOS and Android and add testing and analytics upgrades.
NVIDIA released Nemotron 3 Diarization, an open-weight 100M-parameter model that the company reports ranked #1 on VoiceArena’s Diarization-Bench with a 14.72% DER. The model supports up to eight anonymous speaker channels, streaming and offline modes, and multiple latency points.
Launched this week, Verity Score records what six AI assistants say about a Shopify store and preserves each answer with its sources. The app drafts corrections from a merchant’s text, publishes approved fixes to Shopify, and tracks visits and orders attributed to AI.
OpenAI said it banned ChatGPT accounts that likely originated in Cambodia after they used AI to pose as fake law firms, real attorneys and the FBI’s IC3 to target fraud victims. The operation promoted services via social ads and Telegram and requested upfront fees, including a 15% “service fee.”
V7 converts company files into a Context Graph that agents can query, using GPT-5.6 models and GPT-6 Astra for harder queries. V7 reports accuracy gains, lower costs, and faster, auditable workflows across finance, insurance and asset management.
Authors Arthur Zucker, Simon Brandeis, Luc Georges and Lysandre report that tokenizers v1 preserves v0.23 output while improving encode speed by 3–30× on an Apple M4 Max and scaling at 76% across eight workers. Changes include bitcannon splitting, a thread-local word cache and a no-alloc merge loop.